采用深度级联卷积神经网络的三维点云识别与分割

三维目标识别和模型语义分割在自动驾驶、机器人导航、3D打印和智能交通等领域均有着广泛应用。针对PointNet++未能结合三维模型的上下文几何结构信息的问题,提出一种采用深度级联卷积神经网络的三维点云识别与分割方法。首先,通过构建深度动态图卷积神经网络捕捉点云的深层语义几何特征;其次,通过将深度动态图卷积神经网络作为深度级联卷积神经网络的子网络递归地应用于输入点集的嵌套分区,以充分挖掘三维模型的深层细粒度几何特征;最后,针对点集特征学习中的点云采样不均匀问题,构建一种密度自适应层,利用循环神经网络编码每个采样点的多尺度邻域特征以捕捉上下文细粒度几何特征。实验结果表明,本算法在三维目标识别数据集ModelNet40和MoelNet10上的识别准确率分别为91.9%和94.3%,在语义分割数据集ShapeNet Part,S3DIS和vKITTI上的平均交并比分别为85.6%,58.3%和38.6%。该算法能够提高三维点云目标识别和模型语义分割的准确率,且具有较高的鲁棒性。

Recognition and segmentation of three-dimensional point cloud based on deep cascade convolutional neural network

Three-dimensional(3D)object recognition and model semantic segmentation are widely appliedin fields such as automatic driving,robot navigation,3D printing,and intelligent transportation.With a focuson the inability of PointNet++to integrate contextual geometric structure information,a method for recognition and segmentation of 3D point cloud modes based on a deep cascade Convolutional Neural Network(CNN)was proposed herein.The deep semantic geometric features of the point cloud could be captured via construction of a deep dynamic graph CNN.Subsequently,the deep dynamic graph CNN was applied recursively as a subnetwork of a deep cascade CNN for nested partition of the input point set for full exploration of the fine-grained geometric features of the 3D model.Finally,to address the point cloud sampling nonuniformity problem in point set feature learning,a density adaptive layer was constructed.A recurrent neural network was used to encode the multiscale neighborhood features of each sample point to capture the contextual fine-grained geometric features.The experimental results showed that the recognition accuracy of this algorithm on ModelNet40 and ModelNet10 were 91.9%and 94.3%,respectively.The mean intersection-over-union on the ShapeNet Part,S3DIS,and vKITTI datasets was 85.6%,58.3%,and 38.6%,respectively.This algorithm can improve the accuracy of 3D point cloud recognition and model semantic segmentation,and it shows high robustness.